A DAS-VSP data denoising method

Abstract

This invention relates to the field of well-drilled seismic exploration data processing technology, and provides a method for denoising DAS-VSP data. The method includes: constructing a DAS-VSP dataset under various noise backgrounds; constructing and training a diffusion Transformer network based on a deterministic sampling strategy; and using the trained network to suppress multiple types of noise and recover weak signals from the DAS-VSP data. This invention effectively solves the problems of missing weak signal axes and incomplete noise suppression in existing deep learning models for DAS-VSP data denoising, reducing the loss of effective signal energy in the denoising results. While significantly improving the inference efficiency of the diffusion model, it ensures the fidelity of signal recovery, providing more reliable data support for subsequent high-precision seismic imaging.

Description

Technical Field

[0001] This invention belongs to the field of well seismic exploration data processing technology, and particularly relates to a DAS-VSP data denoising method. Background Technology

[0002] Oil and gas resources play a crucial role in energy supply. As exploration and development shift towards deeper, ultra-deep, and unconventional resources, the difficulty of exploration and identification is increasing, necessitating more advanced geophysical exploration methods. Distributed Acoustic Sensing-Vertical Seismic Profiling (DAS-VSP), utilizing fiber optic cables as seismic signal sensors within the wellbore, offers significant advantages such as full-well-section sampling, resistance to high temperatures and pressures, and high spatiotemporal resolution. It has become a core and commonly used data acquisition method in deep and ultra-deep oil and gas exploration.

[0003] However, during field data acquisition, actual DAS-VSP data processing faces significant challenges due to complex geological environments, wellbore coupling conditions, and external factors such as vibrations of optical components. First, the signal-to-noise ratio (SNR) is extremely low, and the effective signal is interfered with by various complex noises. Second, the noise is extremely complex, including checkerboard noise, horizontal noise, optical noise, and various types of fading noise, which exhibit different characteristics in the time and frequency domains. Furthermore, different noises overlap in the time and frequency domains, making it difficult to suppress them simultaneously using traditional methods and ensuring the fidelity of signal recovery, which can easily lead to signal distortion.

[0004] In recent years, deep learning methods, represented by convolutional neural networks, have been widely applied in the field of seismic data denoising due to their powerful nonlinear mapping capabilities. However, existing deep learning techniques still have some limitations when processing DAS-VSP data: firstly, some networks have an upper limit to their feature extraction capabilities, resulting in the loss of weak signal axes in the denoising results and incomplete noise suppression; secondly, some models often fail to achieve accurate signal-to-noise separation during training, causing some effective wave energy to be mistakenly treated as noise suppression, resulting in serious energy leakage problems. These problems lead to unclear seismic imaging profiles, distorted wavefield information, and consequently, inaccurate velocity model predictions, making it difficult to accurately locate the spatial position and reservoir boundaries of oil and gas reservoirs. Summary of the Invention

[0005] The purpose of this invention is to provide a DAS-VSP data denoising method, which aims to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: a DAS-VSP data denoising method includes the following steps:

[0007] (1) Construct the DAS-VSP dataset for complex geological structures;

[0008] (2) Construct a diffusion model based on Transformer, including:

[0009] Construct a forward diffusion process to obtain diffusion representations for different noise intensities;

[0010] Apply the DDIM sampling strategy to construct a diffusion inverse process;

[0011] A Transformer architecture is constructed to obtain a Transformer-based diffusion model, which is used to estimate the noise distribution at each step and recover a clear target signal from noise pollution in the reverse process.

[0012] (3) Based on the DAS-VSP dataset of complex geological structures, a diffusion model based on Transformer is trained to complete reverse diffusion and recover the target signal.

[0013] This invention combines the strong detail recovery capability of diffusion models with the global information modeling capability of Transformers to better characterize weak signal details through iterative denoising. Unlike traditional diffusion models, this invention uses the DDIM sampling strategy, achieving noise suppression and high-precision signal recovery in just four sampling steps. Furthermore, conditional information guides the sampling process in a defined direction. To address energy leakage, this invention designs a dual-objective optimization strategy, fine-tuning network weights based on signal reconstruction loss. This allows the model to achieve noise suppression while maintaining signal fidelity, improving its amplitude preservation capability. This method can effectively process complex real-world DAS-VSP data and achieves good application results in noise suppression, signal recovery, and signal amplitude preservation. It not only efficiently suppresses complex and diverse DAS-VSP noise but also provides highly reliable data support for velocity model prediction and precise reservoir location in deep oil and gas exploration. This has significant strategic importance for reducing deep exploration risks and improving resource development efficiency. Attached Figure Description

[0014] Figure 1 The horizontal stratigraphic model provided in this embodiment of the invention is as follows: a is a horizontal layer, b is a thin interbedded layer, c is an inclined layer, d is an irregular layer, e is an anticline, and f is a fault.

[0015] Figure 2 This is a diagram illustrating the forward and reverse diffusion processes provided in an embodiment of the present invention.

[0016] Figure 3 This is a general structural diagram of the diffusion Transformer provided in an embodiment of the present invention;

[0017] Figure 4 The diagram shows the structure of the diffusion feedforward network and the diffusion self-attention module provided in the embodiments of the present invention. a is the structure of the diffusion feedforward network and b is the structure of the diffusion self-attention module.

[0018] Figure 5 The following are some one-to-one paired DAS-VSP dataset samples provided in the embodiments of the present invention: a is a clean dataset sample, and b is a noisy dataset sample.

[0019] Figure 6 The simulation data provided in this embodiment of the invention for verifying model performance is as follows: a is the formation model, b is the clean record, c is the noisy record, and d is pure noise.

[0020] Figure 7 for Figure 6 The processing results and difference plot of the noisy record in C, where a is the processing result and b is the difference plot;

[0021] Figure 8 The images show the processing results and difference graphs of actual DAS-VSP data provided in this embodiment of the invention. a represents the actual data, b represents the denoising result, and c represents the difference graph. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0024] Example 1: A DAS-VSP data denoising method, comprising the following steps:

[0025] (1) Constructing the DAS-VSP dataset for complex geological structures:

[0026] Constructing a pure signal set:

[0027] The Ricker wavelet forward modeling method is used to simulate the propagation trajectory of seismic waves, as shown in the following formula:

[0028] ;

[0029] in, For amplitude, For time, Main frequency, Start time;

[0030] Construct a stratigraphic model that includes horizontal layers, thin interbedded layers, inclined layers, anticlines, faults, and irregular layers, such as... Figure 1 As shown in the table (in each model, the red inverted triangle represents the epicenter location, and the black vertical line represents the DAS receiving array in the well; the model scales in the horizontal and depth directions are 1000 m and 2000 m, respectively), corresponding velocity and density parameters are configured for different models. The parameter setting range is shown in Table 1. 7000 data blocks are randomly selected from the generated two-dimensional DAS record using a sliding window with a scale of 256×256:

[0031] Table 1 Forward modeling parameters of stratigraphy

[0032] parameter set up wavelet type Ricker Lane spacing 1m Sampling frequency 2500 Hz clock speed 40-80 Hz density <![CDATA[1500-2350 kg / m 3 ]]> wave velocity 1000-4000 m / s

[0033] Constructing a noisy dataset:

[0034] Multiple noise segments, including horizontal noise, checkerboard noise, optical noise, spike noise, fading noise, and random noise, were extracted from the time window from the first to the last arrival of actual DAS-VSP data. These segments contained no seismic signal components and were considered pure noise segments. These pure noise segments were then superimposed onto the clean signal with different weights, and the same sliding window coordinates were used to extract noisy data blocks, ensuring strict pairing of noisy and clean samples in both spatial location and temporal range. Figure 5 As shown, some noisy and clean signal samples are presented;

[0035] (2) Construct a diffusion model based on Transformer, including:

[0036] Constructing the forward diffusion process:

[0037] Forward process as Figure 2 As shown, following a T-step Markov chain, noise scheduling is used. Controlling the rate of noise addition to pure signals Add Gaussian noise to it, by Time conversion The forward process at time step is defined as follows:

[0038] ;

[0039] in, express arrive The conditional probability distribution, It follows a Gaussian distribution; and for and Diffusion representation at a given moment; , Randomly selected during the forward process This represents the total number of forward diffusion steps; , where is the noise diffusion coefficient; and is the noise timetable used to control the diffusion rate. Depend on arrive Linear increase; Let be the characteristic covariance matrix; the meaning of the entire formula is: the th in the forward noise addition process Step data It is from a source As the mean, with It is obtained by sampling from a Gaussian distribution of the covariance matrix;

[0040] The state transition formulas for the forward process are as follows:

[0041] ;

[0042] in, for The Gaussian noise added at each time point is used to obtain the result through a reparameterization technique. Diffusion representation to any time State transition formula:

[0043] ;

[0044] in, This represents the superposition of Gaussian noise and a pure signal. , where is the cumulative coefficient for different time steps, and is the noise. The expression follows a mean of 0 and a variance of . Gaussian distribution, and and Having the same dimensions;

[0045] The purpose of the forward process is to obtain the diffusion representation of different noise intensities, which is used to train the noise estimation network so that it has the ability to accurately predict noise.

[0046] Constructing the reverse diffusion process:

[0047] Reverse process as Figure 2 As shown, with a pure Gaussian distribution Starting with this, an effective signal is gradually generated through iterative denoising, embedding conditional information within it. , For noisy DAS-VSP data, the two are concatenated in the channel direction to form dual-channel input data. The formula for the reverse process is defined as follows:

[0048] ;

[0049] in, The noise prediction results of the noise estimation network. It is by With estimated noise The weighted average is obtained;

[0050] To reduce the number of sampling steps and sampling time required for the reverse process, the DDIM sampling strategy is applied, utilizing the forward diffusion steps. To obtain a uniform subsequence, the calculation method is as follows:

[0051] ;

[0052] in, For the preset number of sampling steps, This represents the index of the new time step in the "accelerated sampling" process, as described in this embodiment of the invention. This method reduces the number of sampling steps to 4 and shortens the sampling time from 2.62 hours to 34 seconds;

[0053] Building the Transformer architecture:

[0054] The overall structure of the diffusion Transformer is as follows: Figure 3 As shown, the overall structure includes a diffusion process and a Transformer architecture. This Transformer is based on the Restormer design. The encoder and decoder consist of several DTM modules, which include a Diffusion Self-Attention Module (DSA) and a Diffusion Feedforward Network (DFFN). Their structures are as follows: Figure 4 As shown;

[0055] DFFN uses 1×1 pointwise convolution and temporal embedding. The process of obtaining diffusion representations at different scales is as follows:

[0056] ;

[0057] ;

[0058] in, Representation layer normalization; Indicates time step Sine coding mapping; This represents the input tensor data. and They represent Mean and variance in the channel direction, and These are two learnable parameters; Represents a 1×1 pointwise convolution. It is a Gaussian linear unit (GELU). It is by and spliced ​​dual-channel input data;

[0059] The goal of DSA is to learn global information from DAS data using a self-attention mechanism, as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] in, It is a 3×3 depth convolutional layer. These are the feature matrices used for attention coefficient calculation. It is a learnable scaling parameter used to control... and The magnitude of the dot product should be adjusted to prevent the gradient from entering the gradient saturation region of the softmax function. It is a separation operation;

[0064] The Transformer encoder and decoder have a depth of 4, and the number of DTMs at each level from top to bottom is... The number of attention heads in DSA are respectively The number of input data channels is 2, and the number of output data channels is 1;

[0065] The total number of iterations in the dual-objective fine-tuning training phase is set to 90K, and the learning rate scheduling follows the cosine annealing strategy from the first phase. The learning rate will be adjusted from the previous iteration interval. Gradually decrease to The rules for adjusting data size and batch size are the same as in the first stage, but the update frequency is changed to once every 5K iterations.

[0066] (3) Based on the DAS-VSP dataset of complex geological structures, a Transformer-based diffusion model is trained to complete reverse diffusion and recover the target signal:

[0067] Pairing data Its spatial dimensions are These correspond to the two dimensions of time sampling and spatial sampling, respectively, using pure signals. Starting from the time step, randomly select a time step. And obtain the diffusion representation at that moment. Then compare it with the noisy sample Dual-channel input data is obtained by concatenating data along the channel dimension. , and use it for training;

[0068] Transformer output noise residual and with Addition calculation of prediction noise ;

[0069] With actual added Gaussian noise As a monitoring signal, by minimizing and Between The loss optimizes the network weights, enabling the model to perform backdivide and recover the target signal during the testing phase. Loss refers to the mean absolute error (MAE), which is calculated as follows:

[0070] ;

[0071] in, The total number of samples, For the true value, This is a predicted value;

[0072] The optimization objectives of this invention are as follows:

[0073] ;

[0074] Among them, time step , For a pure signal, The actual Gaussian noise added. Indicates the solution loss;

[0075] Since the forward and backward processes are two independent processes, and training only includes Gaussian noise prediction and not signal sampling, this embodiment of the invention introduces a signal reconstruction loss on the basis of the first stage of training to further fine-tune the network weights, making the network pay additional attention to the signal fidelity problem. The reconstruction loss is defined as follows:

[0076] ;

[0077] in, This is the backsampling result of the network weights in the first stage, i.e., the denoising result;

[0078] Through bi-objective optimization, the network can suppress noise while maintaining signal fidelity, effectively mitigating the signal energy leakage problem.

[0079] The actual training and testing process is as follows:

[0080] Model training: Training was performed on a server equipped with an Intel Core i9-13900KF CPU and two NVIDIA GeForce RTX 3090 GPUs. Training employed a dual-GPU parallel approach, with a total of 270K iterations. During training, the input data size and batch size were adjusted every 10K iterations, following the following adjustment rules: The data is calculated in single-precision floating-point (float 32) format, and the optimizer selected is AdamW. In this optimizer... , (here) The momentum parameter of the optimizer, and the noise scheduling in the diffusion process. (Different meanings), initial learning rate is And through the learning rate decay strategy of cosine annealing Gradually reduce to within the range The training and sampling algorithms are shown in Tables 2 and 3:

[0081] Table 2 Algorithm Design in the Training Phase

[0082]

[0083] Table 3 Algorithm Design for Sampling Phase

[0084]

[0085] Data testing: First, a pure DAS record is constructed using elastic wave forward modeling, such as... Figure 6 As shown in Figure a, the velocity model includes eight horizontal layers, with velocities from top to bottom of 1500 m / s, 1750 m / s, 2000 m / s, 2250 m / s, 2550 m / s, 2900 m / s, 3350 m / s, and 3600 m / s. The model has a horizontal range of 1000 m and a depth range of 2000 m, with a trace spacing of 1 m. The black vertical lines in the figure represent the receiver array positions, and the red inverted triangles represent the source positions. The shot-receiver distance is approximately 160 m, the dominant frequency of the excitation wavelet is 40 Hz, and the sampling frequency is 2500 Hz. The resulting clean record is shown below. Figure 6 As shown in Figure b, the noise component consists of various noise segments extracted from the actual recording, such as... Figure 6 As shown in Figure d, the red arrows in the figure mark spike noise, horizontal noise, and optical noise, respectively. These noises are superimposed on the clean record according to their occurrence time in the two-dimensional record to obtain noisy test data. The initial signal-to-noise ratio of this data is -1.713 dB. Figure 6 As shown in c;

[0086] The results of processing the simulation data are as follows Figure 7As shown, Figure 7 Figure 'a' shows the noise suppression result of this embodiment of the invention. In the result, all types of noise are effectively removed, and the effective signal masked by optical noise is also fully recovered, with the signal-to-noise ratio improved to 28.925 dB. Figure 7 In the figure, b represents the difference record of the denoising results. The interpolation record basically does not contain any effective signal axes, which proves that the method of the present invention has a good signal amplitude preservation capability.

[0087] The actual data selected is DAS-VSP data from the Tarim Basin in China, such as... Figure 8 As shown in Figure a, the data contains 1700 seismic traces with a recording duration of 2000 ms. The processing results and difference plot are shown below. Figure 8 As shown in Figure bc, the results show that the method provided by the embodiments of the present invention still exhibits excellent noise suppression performance in actual DAS-VSP. All noise types are effectively suppressed, and deep weak signals are also fully recovered. Only a small amount of signal components remain in the difference graph, which is within the acceptable range for industrial applications.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A DAS-VSP data denoising method, characterized in that, Includes the following steps: (1) Construct the DAS-VSP dataset for complex geological structures; (2) Construct a diffusion model based on Transformer, including: Construct a forward diffusion process to obtain diffusion representations for different noise intensities; Apply the DDIM sampling strategy to construct a diffusion inverse process; A Transformer architecture is constructed to obtain a Transformer-based diffusion model, which is used to estimate the noise distribution at each step and recover a clear target signal from noise pollution in the reverse process. (3) Based on the DAS-VSP dataset of complex geological structures, a diffusion model based on Transformer is trained to complete reverse diffusion and recover the target signal.

2. The DAS-VSP data denoising method according to claim 1, characterized in that, The steps for constructing the DAS-VSP dataset for complex geological structures include: Constructing a pure signal set: The Ricker wavelet forward modeling method is used to simulate the propagation trajectory of seismic waves, as shown in the following formula: ； in, For amplitude, For time, Main frequency, Start time; A stratigraphic model containing horizontal layers, thin interbedded layers, inclined layers, anticlines, faults, and irregular layers was constructed. Data blocks were randomly extracted from the generated two-dimensional DAS record using a sliding window to obtain a clean signal. Constructing a noisy dataset: Multiple noise segments were extracted from the time window from the first arrival to the last arrival of the actual DAS-VSP data. These segments included horizontal noise, checkerboard noise, optical noise, spike noise, fading noise, and random noise. The segments did not contain seismic signal components and were pure noise segments. The pure noise segments were superimposed on the clean signal with different weights, and the noisy data blocks were extracted using the same sliding window coordinates to ensure that the noisy and clean samples were paired in terms of spatial location and time range.

3. The DAS-VSP data denoising method according to claim 1, characterized in that, The steps for constructing the forward diffusion process and obtaining diffusion representations for different noise intensities specifically include: Following a T-step Markov chain, using noise scheduling Controlling the rate of noise addition to pure signals Add Gaussian noise to it, by Time conversion The forward process at time step is defined as follows: ； in, express arrive The conditional probability distribution, It follows a Gaussian distribution; and for and Diffusion representation at a given moment; , Randomly selected during the forward process This represents the total number of forward diffusion steps; , where is the noise diffusion coefficient; and is the noise timetable used to control the diffusion rate. Depend on arrive Linear increase; The characteristic covariance matrix; the first in the forward noise addition process. Step data It is from a source As the mean, with It is obtained by sampling from a Gaussian distribution of the covariance matrix; The state transition formulas for the forward process are as follows: ； in, for The Gaussian noise added at each time point is used to obtain the result through a reparameterization technique. Diffusion representation to any time State transition formula: ； in, This represents the superposition of Gaussian noise and a pure signal. , where is the cumulative coefficient for different time steps, and is the noise. The expression follows a mean of 0 and a variance of . Gaussian distribution, and and They have the same dimensions.

4. The DAS-VSP data denoising method according to claim 3, characterized in that, The steps for constructing the diffusion reverse process using the DDIM sampling strategy include: With pure Gaussian distribution Starting with this, an effective signal is gradually generated through iterative denoising, embedding conditional information within it. , For noisy DAS-VSP data, the two are concatenated in the channel direction to form dual-channel input data. The formula for the reverse process is defined as follows: ； in, The noise prediction results of the noise estimation network. It is by With estimated noise The weighted average is obtained; Applying the DDIM sampling strategy, utilizing the number of forward diffusion steps To obtain a uniform subsequence, the calculation method is as follows: ； in, For the pre-set number of sampling steps, This indicates the index of the new time step.

5. The DAS-VSP data denoising method according to claim 3, characterized in that, The encoder and decoder of the Transformer-based diffusion model consist of several DTM modules, which include DSA and DFFN. DFFN uses 1×1 pointwise convolution and temporal embedding. The process of obtaining diffusion representations at different scales is as follows: ； ； in, Representation layer normalization; Indicates time step Sine coding mapping; This represents the input tensor data. and They represent Mean and variance in the channel direction, and These are two learnable parameters; Represents a 1×1 pointwise convolution. It is a Gaussian linear unit. It is by and spliced ​​dual-channel input data; The DSA process is as follows: ； ； ； in, It is a 3×3 depth convolutional layer. These are the feature matrices used for attention coefficient calculation. It is a learnable scaling parameter used to control... and The magnitude of the dot product should be adjusted to prevent the gradient from entering the gradient saturation region of the softmax function. It is a separation operation.

6. The DAS-VSP data denoising method according to claim 2, characterized in that, Based on the DAS-VSP dataset of complex geological structures, a Transformer-based diffusion model is trained to complete the steps of backdiffusion and recovery of the target signal, specifically including: Pairing data Its spatial dimensions are These correspond to the two dimensions of time sampling and spatial sampling, respectively, using pure signals. Starting from the time step, randomly select a time step. and obtain time steps diffusion representation Then compare it with the noisy sample Dual-channel input data is obtained by concatenating data along the channel dimension. , and use it for training; Transformer output noise residual and with Addition calculation of prediction noise ; With actual added Gaussian noise As a monitoring signal, by minimizing and Between Loss optimization network weights The loss is calculated as follows: ； in, The total number of samples, For the true value, This is a predicted value; The optimization goals are as follows: ； Among them, time step , For a pure signal, The actual Gaussian noise added. Indicates the solution Loss; the forward and backward processes are two independent processes. Training only includes Gaussian noise prediction and does not include signal sampling. A signal reconstruction loss is introduced to fine-tune the network weights. The reconstruction loss is defined as follows: ； in, This is the backsampling result of the network weights in the first stage, i.e., the denoising result.

Images